Creative Artificially-Intelligent Agents for the Arts: An Interdisciplinary Science-and-Arts Approach

Coordinator: Jonas Braasch
Advisors:
Selmer Bringsjord, USA
Ted Krueger, USA
Johannes Goebel, USA
Pauline Oliveros, USA

Artificial Intelligence (AI) has made impressive progress since its start in 1956.  It now influences our daily lives, as AI systems are an integral part of consumer technology today, from SIRI to automobiles to Semantic Web.  However, while AI systems can be very successful if they are precisely told what to do (e.g., perform a parallel parking task, play chess), they are usually useless if the objectives are not clearly spelled out.  They can learn along a precisely given trajectory (e.g., to learn to understand spoken text or compose an instrumental music piece in the tradition of JS Bach), but they don’t break rules to produce something more exciting.  Deep Blue can play chess, but if you present it with a game implemented on a chess board, it will be lost.  In short, machines are simply not very creative.

The idea of this white paper is to form an intellectual think tank to overcome existing roadblocks and investigate alternative strategies in AI.  Among the items that will be discussed is the implementation of design oriented processes for AI systems.  Artists and designers often work on a less hypothesis- or goal-driven approach as compared to scientists and engineers; they pursue an open-ended, purely experimental approach instead, where the outcome of each phase informs the next one, not necessarily having a fixed goal in mind.  Along with this approach, there is a need for better AI evaluation systems that can judge the outcome more freely than just examining the results along an externally given set of rules.  Using the experience of artists with the abstract, can we make agents more creative by allowing them to be continuously evaluate what they accomplish?  How can we create AI systems that can develop and evaluate their own concepts?

Part of this discussion will include the creation of a network for more complex AI systems that simulate several areas of the brain, or the abstract AI equivalent, simultaneously, by using a meta-concept to connect existing AI modules using a UDP protocol in a computer-cluster network.  Another central aspect are systems that can draw on different algorithms to perform a task, making the selection part of the creative process.  Along the same lines, we can look into web data-mining methods that allow these machines to receive information beyond what is given to them by the experimenter.